Image-to-video prompting: camera motion without identity drift

Motion prompts are not image prompts

The image already defines appearance. Re-describing every visual detail can invite the video model to reinterpret the frame. Focus on continuity, camera behavior, subject action, environmental motion, and prohibited changes.

A compact motion prompt structure

A useful prompt can be short when each instruction has a clear job.

• Frame lock: preserve the uploaded image as the exact first frame.
• Subject lock: keep identity, product, body, vehicle, or wardrobe stable.
• Camera: dolly, slider, orbit, crane, handheld, or locked tripod.
• Action: one readable movement with realistic speed and weight.
• Secondary motion: light, fabric, hair, reflections, smoke, or water.
• Prohibitions: no morphing, duplicate objects, new text, or camera shake.

Split motion into four production rows

Before writing the final instruction, separate camera, subject, environment, and edit. The camera row defines path, speed, height, distance, and endpoint. The subject row contains one readable performance or product action. The environment row contains only necessary secondary movement. The edit row records duration, usable entry and exit, transition, and sound cue.

If the camera and subject both demand attention, simplify one or divide the moment into separate shots. This Motion Split makes the brief portable across tools and makes a failed generation easier to diagnose than a paragraph where every movement competes at once.

Treat a motion reference as a second source of truth

A motion-reference workflow combines an appearance source with a movement source. Confirm that the motion clip is owned, commissioned, licensed, or otherwise permitted. Then compare framing, body visibility, clothing, environment, and camera behavior with the approved character frame before generation.

A movement source with a moving camera or hidden limbs may force the model to invent information that the still does not contain. Scene modes and controls can change between product versions, so review the complete clip rather than treating a selected mode as a consistency guarantee.

Edit around model limits

Not every shot needs complex action. Controlled detail shots, slow pushes, and short inserts often create a more expensive sequence than forcing one generation to perform an entire scene.

Write the motion instruction in layers

Begin with continuity: preserve the exact subject, product, vehicle, wardrobe, environment, composition, and text already approved in the frame. Add one camera instruction with direction and speed. Add one dominant subject action. Then describe limited secondary movement such as hair, fabric, reflections, smoke, water, or light.

End with a short prohibition block based on the actual risk: no morphing, duplicate objects, new text, redesigned product details, unstable hands, background replacement, or unrequested camera shake.

Plan duration around motion complexity

Longer generations create more time for identity and geometry to drift. Use the shortest duration that completes the shot’s editorial job. A detail insert may need only a subtle push; a reveal needs enough time for the camera move to read but not enough to invent a second scene.

When a shot needs several actions, split it into separate approved source frames and edit the results. Do not force a single generation to perform a complete sequence.

Build continuity between separate shots

Record screen direction, camera height, lens feel, light direction, subject orientation, prop position, and the intended cut. The end of one shot and the first frame of the next do not need to be identical, but they must agree on the visual facts the viewer uses to understand space and identity.

Judge temporal stability, physical weight, edge consistency, reflections, hands, labels, and background motion—not only whether the camera moved.

Diagnostic table

Find the failed layer before regenerating.